SAM Fine-Tuned on Synthetic Data for Mitochondria Segmentation
Researchers propose fine-tuning the Segment Anything Model (SAM) on synthetically generated fluorescence microscopy data to overcome domain shift and data scarcity for mitochondria instance segmentation. The method simulates realistic mitochondria and emulates optical properties of fluorescence microscopes to create large-scale annotated datasets. The fine-tuned model is evaluated on curated data, aiming to improve morphological analysis critical for understanding cellular health, energy production, and metabolic regulation.
Key facts
- SAM is fine-tuned exclusively on synthetically generated FM data.
- Synthetic data simulates realistic mitochondria and emulates optical properties of fluorescence microscopes.
- Domain shift from natural images to FM includes diffraction-limited resolution, low contrast, and overlapping organelle networks.
- Severe lack of high-quality manually annotated instance segmentation datasets for mitochondria.
- Morphological analysis of mitochondria in FM is crucial for cellular health, energy production, and metabolic regulation.
- The paper is published on arXiv with ID 2605.31284v1.
- The approach provides a scalable solution to data scarcity.
- Evaluation is performed on a curated dataset.
Entities
Institutions
- arXiv